Ab Initio Versions ((new)) May 2026

We talk a lot about machine learning potentials, DFT surrogates, and foundation models for materials. But here’s a quiet truth: every new, truly predictive method still starts with an ab initio version.

And if you’re building something new — start with the ab initio version. Even if it only runs on 10 atoms. ab initio versions

ML potentials are getting shockingly good. But they depend on training data — and that data comes from the expensive, “ab initio version” codes. When the ab initio version changes (e.g., higher accuracy functional, core-valence correlation), the ML model’s ceiling moves too. We talk a lot about machine learning potentials,

Here’s a draft for an interesting post about ab initio versions — tailored for a computational chemistry, materials science, or ML/physics audience. Why “Ab Initio” Versions Still Matter in an AI-Driven World Even if it only runs on 10 atoms

The first implementation of a theory — no experimental fitting, no empirical parameters, just fundamental constants and equations.